Geographic Concentration in the U.S. Retail and Wholesale Sectors

manufacturing data for 1987 and ask the question: “How concentrated are manufacturing industries?” Using a model of location choice that incorporates spillovers,the authors derive measur

Preliminary Results – Please Do Not Cite

Geographic Concentration in the U.S Retail and Wholesale Sectors

*Center for Economic Studies, U.S Bureau of the Census, 4700 Silver Hill Road Stop

6300, Washington Plaza II Suite 211, Washington D.C 20233-6300

(sklimek@ces.census.gov)

**5000 Forbes Avenue, Carnegie-Mellon University, H John Heinz III School of Public Policy and Management, Census Research Data Center, Room 238, Pittsburgh, PA 15213 (dmerrell@andrew.cmu.edu)

I Introduction

When looking at the economic landscape, one can point to a number of instances where it appears that industries locate in the same geographic regions High technology industries seem to be located in Silicon Valley, automobiles in Detroit, financial

industries in New York and Chicago, and tires in Dayton—to give a few examples In thepast, these stories of geographic industrial concentration have been taken as a rule of thumb and were not given a lot of attention in the economics literature since Marshall (1920)

More recently, however, interest has rekindled on the subject, and a good deal of attention has focused on the geographic concentration of industries Krugman (1991) makes the case that this sort of concentration may be the general rule rather than an exception—that the agglomeration of industries is more than merely a set of examples that one can point out such as Silicon Valley Rather, this sort of agglomeration of

industries could be the source of the increasing returns that models of international trade and economic growth have at their core

This renewed interest in the geographic concentration of industries seeks to explain why agglomeration exists in the first place Some of this attention focuses on the role that technological spillovers play in industry concentration Glaeser, Kallal, Scheinkman,and Schleifer (1992) examine the role of technological spillovers in the growth of cities They find evidence that spillovers between industries may be more important than

spillovers within an industry They also find evidence that competition and economic diversity supports employment growth while specialization damages growth Jaffe, Trajtenberg, and Henderson (1993) provide evidence of technological spillovers using

patent citation data This work also finds evidence of the geographic localization of spillovers and finds further that these spillovers can be quite large and significant.

Other recent work takes a step back to examine not why these sorts of

agglomerations exist, but rather whether or not they are real Ellison and Glaeser (1997) examine U.S manufacturing data for 1987 and ask the question: “How concentrated are manufacturing industries?” Using a model of location choice that incorporates spillovers,the authors derive measures of geographic concentration These indices control both for the size distribution of manufacturing plants and for the size distribution of the

geographic areas Using these indices, the authors find that nearly all industries in their study exhibit some degree of geographic concentration Additionally, the authors find that there is evidence of co-agglomeration of industries in the sense that there appears to

be location choices hinging on the upstream-downstream relationships among businesses

In related work, Dumais, Ellison, and Glaeser (1997) treat the location of

manufacturing plants as a dynamic process The idea is that the entry, exit, expansion, and contraction of manufacturing plants will also affect measures of geographic industrialconcentration Using the Census Bureau’s Longitudinal Research Database, these authorsfind that the location choices of new plants and the differentials in the growth rates of plants tend to reduce levels of geographic concentration in U.S manufacturing

Additionally, they find that the exit of plants tends to increase agglomeration.1

Virtually all of the research focusing on the agglomeration of industries uses data

on the manufacturing sector In this paper, we extend the literature on geographic

concentration by measuring the agglomeration and co-agglomeration in the U.S

1 Maurell and Sedillot (1999) use indices similar to Ellison and Glaeser (1997) to examine the geographic concentration of French manufacturing plants These authors find that patterns of geographic concentration

in France are very similar to the patterns in the U.S data.

wholesale and retail sectors in 1992 Together, these two sectors account for about as much economic activity as manufacturing In 1997, the wholesale and retain sectors accounted for about 15% of all economic activity in the United States; in the same year, manufacturing accounted for roughly 17% of economic activity The shear size of the wholesale and retail sectors makes them important to understand Using a newly

constructed dataset containing the statistical universe of wholesale and retail

establishments, we use the Ellison-Glaeser geographic concentration indices to examine patterns of agglomeration within and across the industries in these two sectors—using calculations from the manufacturing sector as a benchmark.2

Our contributions are threefold First, little is know about geographic concentrationoutside of the manufacturing sector In the spirit of Ellison and Glaeser, this paper simplytakes a step back to answer the primitive question: do establishments in the U.S

wholesale and retail sectors tend to be geographically concentrated? We use the same measures in manufacturing as a benchmark Second, we focus on the co-agglomeration

of industries in retail and wholesale To do this we calculate measures of

co-agglomeration for all possible combinations of industries and examine the distribution of these measures Third, the foundation of the U.S statistical program has been the

Standard Industrial Classification (SIC) system However, after 1997 all economic census data will be collected under the new North American Industrial Classification System (NAICS) The conversion to NAICS represents a significant change in the way economic census data are collected and reported This paper uses data from the 1992

2 These data are maintained at the Census Bureau’s Center for Economic Studies and are collected under the Census of Wholesale Trade and the Census of Retail Trade programs.

Economic Census that have been converted from SIC to NAICS and provides an

introduction to the new NAICS sectors, subsectors, and industries

In Section II, we detail the model used by Ellison and Glaeser and the measure of geographic concentration and co-agglomeration that is derived from it Section III describes the data Section IV contrasts the results for manufacturing to retail and

wholesale and details the analysis of co-agglomeration Section V concludes

II The Model

There likely are many reasons why an industry may be geographically concentrated, but two broad motivations spring to mind First, depending on the type of industry, some locations may present certain natural advantages over others For example, industries requiring large amounts of warehousing may locate near commercial naval ports or near major highways Second, certain industries may tend to be concentrated because of technological spillovers that accrue; this certainly could explain the location of the computer industry in Silicon Valley However, in general, it likely is the case that the concentration of industries is some amalgam of both of these broad motivations Glaeser and Ellison (1997) develop a model of location choice that incorporates both of these motivations, and from that model, indices of geographic concentration are derived We present a shortened version of their model below

( k ) ki i

i

π =log + 1, , −1 +log

where logπki is profit accruing to business unit k located in region i Business unit k’s

profits are a function of logπi , the profit of the “typical” firm located in region i; this

profit also is a function of observable and unobservable regional characteristics Profit also is a function of g i(υ1, ,υk−1) the effects of spillovers from the other k-1 business units located in region i Finally, profit is a function of an idiosyncratic shock for

business unit k, located in region i We assume that that the { }εki are independent Weibullrandom variables that also are independent of the { }πi Further, it should be clear that if

( 1, , k−1) ≡0

i

g υ υ ∀ i, then the model reduces to a standard conditional logistic model—

conditioned on the realizations of the { }πi Next, we impose the following two

parametric restrictions on the model:

j j

other regions In the latter case (viz., γna =1), all k business units would find an optimum

by locating in that region

Conditions (1) and (2) incorporate natural advantages into the location decisions

of businesses Next is the incorporation of technological spillovers into the location decision calculus The idea is that locating near other facilities in the same industry couldrepresent lower transportation costs or even the transfer of knowledge across facilities Section II.B details how we incorporate spillovers

B Spillovers

Consider the following model of plant location that incorporates spillovers

ki k

l

il kl i

≠

))(

1(log

log

The {e ki} are Bernoulli random variables equal to one with probability γs∈[0,1], and equal to zero with probability 1-γs The variable u il is a dichotomous indicator variable

equal to one if establishment l is in region i, or equal to zero otherwise The importance

of spillovers is captured by the probability parameter γs

C Measures of Geographic Concentration and Co-agglomeration

For a single industry with M geographic regions (counties) and N business units

(establishments), Ellison and Glaeser use γ as the measure of geographic concentration, where γ is defined as:

i i

N j j M

i i M

i

i i

i

i

i i

z x

z x

x s H

x

H x G

1

2 1

2

1 2 2

1

2 1

2

2

2

1 1

1 ) (

) 1 ( 1

1

γ

where si is the share of industry employment in region i, xi is the share of total

employment in region i, and zi is the share of establishment employment of the industry.Ellison and Glaeser show that if the models in section IIA and IIB describe plants’ location decisions, then the measure of geographic concentration is a useful measure that has several desirable properties First, the index can be calculated easily with the

information in our dataset Access to the establishment level data means that each component of the index can be calculated by aggregating up to the county or industry level Second, the scale of the index allows comparisons to be made to a benchmark of

“no-agglomeration” when the expected value of γ is equal to zero Third, the index is comparable across industries in which the size distribution of firms differs

Ellison and Glaeser extend the model in Section II to examine the extent to which industries are co-agglomerated The measure γc, defined below, shares the desirable properties of γ; to be sure, γc and γ share the same scale

r j

j j

j i

i c

w

H w

H x

G

1 2

1

2 2

1

) 1

( ˆ 1

γ γ

where G=∑(s i −x i)2is area i's share of the aggregate employment in the r industries

and H =∑j w2j H j

is an establishment’s Herfindahl index of the aggregate of the r industries

If the measure γc is equal to zero, then there are no spillovers or natural advantages specific to the industry group; rather the natural advantages would be specific to an industry—not the industry group An alternative measure of co-agglomeration is λ:

∑

=

j

j j

c

γ λ

ˆ

If the measure λ is close to one, then all spillovers and natural advantages are group specific rather than industry specific On the other hand, a value of λ close to zero implies that the industries exhibit little co-agglomeration—that spillovers and natural advantages are industry specific and not group specific We use the measure λ to analyze

to extent to which industries are co-agglomerated.3

C The Data

Our data come from two sources First, we use establishment level data from the

1992 Economic Census An establishment level observation is defined as a business or

an industrial unit located at a single physical location Further, all establishments must produce goods, distribute goods, or perform services.4 The Economic Census covers the universe of retail and wholesale establishments in the United States These observations contain a wealth of information on the employment, sales, wages, industry and

geographic characteristics, inter alia, of business units Second, to get information on

regional characteristics, we use data from Counties 1996 From this second source of

information, we construct total county employment for 1992 These two sources of data

3 To be sure, λ measures the strength of co-agglomerative forces relative to agglomerative forces.

4 This definition is not always correct The Census Bureau sometimes splits up very large "establishments" into several establishments, especially when the products and industries these plants produce in are quite varied In our analysis, we do not exclude establishments that are broken out in this way.

provide the information necessary to construct the Ellison-Glaeser geographic

concentration indices

This paper uses data on wholesale and retail establishments in the United States thatare classified using the new North American Industrial Classification System (NAICS); all other work of which we are aware uses the Standard Industrial Classification (SIC) system to define an industries There are substantial differences between the NAICS and SIC systems In the following paragraph, we give a brief overview of the NAICS

taxonomy

A NAICS subsector is the three-digit code—comparable to the two-digit SIC code There are two more detailed breakdowns, the five-digit NAICS code with is referred to asthe NAICS industry, and the six-digit NAICS code which is referred to as the U.S

industry The combination of NAICS industries and U.S industries is comparable to the old four-digit SIC industries Under the SIC system, there were 459 four-digit industries

in manufacturing, under NAICS that number increased to the 474 NAICS leaves the number of wholesale industries constant at 69, and increases the number of retail

industries from 64 to 72 More importantly, NAICS redefines the boundary between the two sectors, which results in a number of establishments moving from wholesale to retail.Klimek and Merrell (1999) provide a very detailed discussion on the differences between the NAICS and SIC industry classification taxonomies

For the analysis that follows, we compute our indices of geographic concentration using the NAICS and U.S industries—keeping mind that these levels of aggregation correspond to the four-digit SIC levels For our indices of co-agglomeration between

wholesale and retail establishments, we use the sub-sector code as the industry group of analysis.

VI Geographic Concentration Results

First, we describe the general results for the manufacturing sector Assuming the null hypothesis of γs=γna=0 (viz., that there are no spillovers or natural advantages that would give rise to geographic concentration), we compare the raw concentration

= (s i x i)2

G to the expected value of G under the null.5 We find a comparable, but weaker result than Ellison and Glaeser using data for 1992 Of the 469 manufacturing industries, 433 industries have a value of G that is greater than the expected value of G This implies that 433 (92.3%) manufacturing industries are more geographically

concentrated that what would be expected to arise if establishments were located

randomly In contrast, 36 (7.7%) industries are more evenly distributed than what would

be expected if establishments were located randomly Calculating the variance of G, we check to see if the difference is significant.6 Of the 433 manufacturing industries where the difference between G and E[G] is positive, only 12.2% are significantly different which suggests that there is not a lot more geographic concentration than one might expect

to arise of plants were located randomly This is in contrast to the 82.7% found by Ellison and Glaeser using aggregate 1987 data For the 36 manufacturing industries where the difference is negative, none are significant—a result similar to Ellison and Glaeser

The results for retail and wholesale are strikingly similar Of the 130 retail and wholesale industries, we find that the difference is positive for 128 industries (98.5%), and

1 1

2 0

2

j j

j r j

r j j

x G

=

=

γγ

6 The formula for var(G) is described on page 907 of Ellison and Glaeser (1997).